Vaultsmith

AI for Obsidian that works like the vault owner, not like a generic note bot.

Vaultsmith is a local agent launcher and review-first workflow for Obsidian vaults. It sets up vault-local instructions, starts codex or claude inside tmux, and gives the agent a constrained operating surface so note creation and organization can stay consistent with how the vault already works.

Why Vaultsmith exists

Most AI note tooling can generate text.

What it usually does badly:

• titles notes in a way that does not match the vault
• picks the wrong folder
• ignores linking habits and structure conventions
• rewrites too much when a small edit would do
• makes users manually clean up the result afterward

Vaultsmith exists to push the agent in the opposite direction:

• inspect the vault before making style claims
• prefer minimal edits over broad rewrites
• keep proposals short and reviewable
• treat vault conventions as a first-class constraint
• make approval explicit before applying edits

What it does today

Vaultsmith currently ships a practical local workflow:

• initializes a vault-local .vaultsmith/ workspace
• writes role prompts and bootstrap instructions for Codex and Claude
• launches agent sessions in tmux
• defaults to a review-first flow with proposal and receipt files
• lets you inspect status, tail output, approve, and apply
• keeps a small vault memory file for durable conventions

It is intentionally narrow right now. The current product is a launcher and guardrail layer for agent-driven vault work, not a full autonomous vault intelligence system yet.

Core idea

Vaultsmith should help an agent behave more like the person who owns the vault.

That means the agent should learn and respect signals such as:

• naming patterns
• folder placement habits
• heading structure
• bullet and checklist style
• link and tag behavior
• frontmatter usage
• note family conventions

The long-term direction is simple: when Vaultsmith creates or reorganizes a note, it should feel native to the vault immediately.

Quick start

Requirements:

• Python 3.11+
• pipx
• tmux
• at least one supported agent CLI on PATH: codex or claude

One-line install path

Today, the shortest install path is directly from GitHub:

pipx install git+https://github.com/ronut01/Vaultsmith.git
vsm setup ~/Obsidian/MyVault
cd ~/Obsidian/MyVault
vsm run -- "정리해줘"

Why pipx is the default

Vaultsmith is a CLI tool, not a library. pipx is the cleanest default because it installs the command into an isolated environment without making users create a project virtualenv just to try the tool.

If you prefer pip, install Vaultsmith into a virtual environment:

python -m venv .venv
source .venv/bin/activate
python -m pip install git+https://github.com/ronut01/Vaultsmith.git

Install for local development

If you are working on Vaultsmith itself:

git clone https://github.com/ronut01/Vaultsmith.git
cd Vaultsmith
python -m venv .venv
source .venv/bin/activate
python -m pip install -e . pytest

Distribution status

Vaultsmith is not published to PyPI yet.

The intended release path is:

• pipx install vaultsmith for end users
• python -m pip install vaultsmith inside a virtualenv for users who prefer pip

First commands

Interactive session:

vsm chat

One-shot request:

vsm run -- "현재 정리 안 된 노트 정리해줘"

Review-first workflow

Vaultsmith is designed so the agent proposes work before it applies work.

Typical flow:

vsm run -- "강의 영상 링크 정리해줘"
vsm review
vsm approve
vsm apply

Useful follow-up commands:

vsm status
vsm tail
vsm sessions
vsm resume <session-id>

Behavior notes:

• vsm run attaches by default so you can respond to trust prompts and other interactive agent questions
• vsm run --detach -- "..." keeps the run in the background and prints recent tmux output
• vsm apply does not patch files itself; it re-dispatches the approved session to the underlying agent and then expects a receipt.md

What gets written into a vault

Running vsm setup <vault> creates a small control plane inside the target vault:

.vaultsmith/
  config.toml
  instructions/
    codex.md
    claude.md
  memory/
    vault-summary.md
  roles/
    input-agent.md
    vault-analyst.md
    research-scout.md
    draft-agent.md
    consistency-reviewer.md
  sessions/
    <session-id>/
      session.json
      proposal.md
      approval.md
      receipt.md
      changes.json
AGENTS.md
CLAUDE.md

Session files matter:

• session.json: metadata, mode, runtime state, approval state
• proposal.md: short proposed plan and draft output before edits are applied
• approval.md: approval checkpoint
• receipt.md: short execution summary after apply
• changes.json: proposed operations payload when the session provides one

Command overview

vsm setup <path>          # initialize Vaultsmith in a vault
vsm chat [path]           # start an interactive agent session
vsm run [path] -- "..."   # start a one-shot request
vsm status [session-id]   # inspect current or selected session
vsm tail [session-id]     # show recent tmux output
vsm review [session-id]   # print proposal.md
vsm approve [session-id]  # mark session approved
vsm apply [session-id]    # dispatch approved work
vsm sessions              # list recent sessions
vsm resume <session-id>   # re-attach to tmux session
vsm alias enable vs       # install a shorthand shell alias

Philosophy

Vaultsmith is opinionated in a few ways.

1. Inspect before generating

The vault is the source of truth. The agent should look at local evidence before inventing a structure or style.

2. Small context beats full-vault scans

For most requests, a few relevant notes are better than a noisy global survey.

3. Review before apply

Users should be able to inspect, reject, or refine proposed changes before the agent edits the vault.

4. Minimal edits over dramatic rewrites

Most vault work is maintenance, not greenfield writing. The safe default is a narrow change.

Current scope vs. direction

Current scope:

• local CLI
• tmux-backed agent launch
• vault bootstrap files
• review-first session workflow
• basic session state and approval handling

Planned direction:

• stronger vault-style profiling
• better note placement and naming inference
• more explicit explanation for why a note was titled or placed a certain way
• better support for recurring note families like meetings, research notes, and project updates

Development

Run tests:

cd Vaultsmith
source .venv/bin/activate  # or the virtualenv you use for local development
pytest -q

Smoke test against a temporary vault:

mkdir -p /tmp/vsm-smoke-vault
vsm setup /tmp/vsm-smoke-vault
cd /tmp/vsm-smoke-vault
vsm run -- "테스트 노트를 하나 만들고 지금 vault 상태를 설명해줘"

Status

Vaultsmith is early, but the direction is deliberate.

The goal is not to bolt AI onto Obsidian. The goal is to make an agent operate inside a vault with enough structure, memory, and review pressure that the output starts to feel like it belongs there.